一种利用光子图像解决深度学习中子CT缺乏成对数据的方法  

A method to solve lack of paired data in neutron computed tomography for deep learning by using photon images

作　　者：郭虎 陈帅[2] 杨明翰[2] 张子恒 邵慧[4] 汪建业[2] GUO Hu;CHEN Shuai;YANG Minghan;ZHANG Ziheng;SHAO Hui;WANG Jianye(Institutes of Physical Science and Information Technology,Anhui University,Hefei 230601,China;Institute of Nuclear Energy Safety Technology,HFIPS,Chinese Academy of Sciences,Hefei 230031,China;University of Science and Technology of China,Hefei 230026,China;Anhui Jianzhu University,Hefei 230022,China)

机构地区：[1]安徽大学物质科学与信息技术研究院,安徽合肥230601 [2]中国科学院合肥物质科学研究院核能安全技术研究所,安徽合肥230031 [3]中国科学技术大学,安徽合肥230026 [4]安徽建筑大学,安徽合肥230022

出　　处：《量子电子学报》2024年第5期738-751,共14页Chinese Journal of Quantum Electronics

基　　金：安徽省自然科学基金(2108085QF285);合肥市自然科学基金(2021003);安徽省古建筑智能感知与高维建模国际联合研究中心开放课题(GJZZX2021KF04)。

摘　　要：由于缺少高质量成对数据集,深度学习在中子层析(CT)重建中的应用与发展受到了严重阻碍。虽然中子层析和光子层析成像原理都基于Radon变换,但二者在粒子输运过程中的成像特征不同,所以光子层析图像训练的网络无法直接用于解决中子层析成像的重建问题。为此,本文提出了一种可以解决光子和中子层析成像迁移过程中概率分布差异问题的无监督域适应网络。该方法通过引入最大均值差异计算以减小光子与中子层析图像特征之间的分布差异,并通过小波变换与卷积神经网络相结合的方式增强重建的有效特征。与其他算法的对比验证表明,该方法能够从低投影通量下的中子层析结果中重建出高质量的中子层析图像,有效缓解了低投影通量下中子层析成像退化的问题。Due to the lack of high-quality paired datasets,the application and development of deep learning in neutron computed tomography(CT)reconstruction are severely hindered.Although the imaging principles of neutron CT and photon CT are both based on the Radon transform,the imaging characteristics of the two processes during particle transport are different,so the network trained for photon CT cannot be directly used to solve the reconstruction problem of neutron CT.Therefore,in this work,an unsupervised domain adaptive network is proposed that can solve the probability distribution difference problem in the migration process from photon tomography to neutron tomography.In the proposed method,the maximum mean difference is introduced to reduce the distribution difference between photon and neutron tomography image features,and furthermore,wavelet transform and convolution neural network are combined to enhance the effective features of reconstruction.The comparison experiments with other algorithms show that the proposed method can reconstruct high-quality neutron tomography images from low-flux neutron tomography results,effectively alleviating the degradation of low-flux neutron tomography.

关 键 词：图像处理 中子CT重建 域适应迁移学习 稀疏层析 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]